Abstract: During the intelligent mining process in coal mines, the phenomenon of backwater from the drilling not only affects the safety and efficiency of the drilling operations, but also may cause equipment damage and personnel injuries. Therefore, accurately and efficiently identifying the phenomenon of backwater from the drilling openings is of great significance for ensuring the safety of the operations and improving the automation level of intelligent drilling machines. In view of the problems such as low accuracy, slow response and poor robustness in the current identification of backwater at borehole openings of underground coal mines, an intelligent visual recognition method for backwater at drilling openings of the drilling rig based on the improved YOLOv5 is proposed, which effectively improves the accuracy and real-time performance of the identification. We has established a complete visual recognition system for backwater at drilling openings, including core components such as explosion-proof cameras, AI image processors, and drilling rig controllers. Through self-developed visual recognition devices, the system realizes the up-down adjustment and 360° rotation functions of the cameras, thereby obtaining video images from different angles. This avoids recognition failures caused by complex working conditions such as obstructions or backlighting, significantly improving the adaptability and recognition accuracy of the system. In terms of the model structure, several optimizations were made to the original network structure of YOLOv5. A TRANS (transformer) module was introduced in the backbone network to enhance the feature extraction capability of model; in the neck network, the GhostBottleneck structure was adopted to replace the traditional cross stage partial (CSP) module, which reduced the computational load and improved the lightweighting level of model; in addition, the activation function was replaced from sigmoid linear unit (SiLU) to the more efficient Hardswish, further enhancing the nonlinear expression ability and inference speed of the model. By using the self-collected and labeled dataset of waterback images from drilling openings, the model training and optimization were completed. To verify the effectiveness of the proposed method, the improved YOLOv5 model was compared with YOLOv3, YOLOv4, YOLOv5, YOLOv8n through experimental tests. The results show that the improved YOLOv5 has increased the recognition accuracy by 1.49%, 1.06%, 0.49%, and 0.18% respectively, and the average recognition time has been shortened to 7.4 ms. It demonstrates excellent recognition performance and real-time response capability. Currently, this system has been deployed and applied in 4 coal mines in Inner Mongolia-Shaanxi region. The on-site test results show that the accuracy rate of backwater recognition at drilling openings is as high as 99.63%, which is significantly better than traditional methods. The system operates stably and reliably.